Papers

Diffusion multimodal large language models (dMLLMs) frequently produce long-form outputs marred by semantic drift and repetition, with quality generally degrading as output length increases. We identify two structural deficiencies in existing decoding methods as primary drivers of these failures: confidence-based scoring ignores decoded-neighbor support, and block partitioning prevents access to high-readiness semantic anchors, together causing tokens to be committed before their local context is sufficiently established. We propose \ours{} (Context-Aware Cluster Decoding), a training-free decoding method that scores each masked position by a multiplicative composite of softmax confidence and neighbor proximity, promoting contextually ready tokens above isolated candidates while suppressing low-confidence positional noise, operating block-free to keep high-readiness anchors globally accessible. \ours{} further applies architecture-aware calibration to handle confidence heterogeneity induced by diverse visual integration strategies. Experiments on three dMLLMs across four benchmarks demonstrate consistent quality gains and hallucination reduction over Original, with larger gains in several longer generation settings, highlighting the importance of neighbor support and visual integration strategy for future dMLLM decoding method design. Our code is openly available at https://github.com/zhaoyk-sysu/CACD-dMLLM.

Vision-language models (VLMs) are increasingly being used for document understanding, yet their role in Arabic and Islamic manuscript recognition remains underexplored. To address such a gap in this paper, we evaluate traditional OCR, general-purpose VLMs, Arabic-specialized VLMs, and OCR-conditioned VLM correction across eight Arabic text datasets spanning historical manuscripts, aged printed books, clean print, multi-domain documents, and handwriting. The results show that no single approach dominates across setups. On line-level historical manuscripts, VLMs are close to Tesseract; on page-level manuscript images, they perform better; and in several settings, an OCR-conditioned corrector improves over both standalone OCR and standalone VLMs. The central finding is an OCR-prior recoverability principle: OCR conditioning helps when the OCR output remains visually and textually recoverable, providing anchors that the VLM can refine against the image. It improves recognition on aged print, clean print, mixed-domain Arabic, and some Naskh manuscripts, but degrades performance when the prior is script-mismatched or systematically misleading, as in Maghribi manuscripts and realistic student handwriting. Additional diagnostics show that Arabic VLM-OCR is sensitive to diacritics, preprocessing, generation budget, and repetition loops. These findings support an adaptive OCR-VLM workflow that routes pages according to script, OCR-prior recoverability, length diagnostics, and failure-mode indicators.

World Action Models (WAMs) generate both future video and robot actions, offering two connected outputs for post-training supervision. How can a pretrained WAM learn from a stronger Teacher on the histories it encounters during execution? We present WAM-OPD, which collects Student rollout histories and queries a Teacher for paired video and action targets. The Student learns from both targets while retaining its one-step video and action generation at deployment. Across 12 RoboTwin 2.0 tasks, WAM-OPD improves average success from 33.8% to 65.7%; across four real-robot tasks, it improves average success from 51.4% to 64.6%. With the collected Student histories held fixed, joint video-action supervision achieves the highest observed success on all three ablation tasks, while either modality alone also improves performance. A separate comparison with Teacher-generated histories finds task-dependent differences between the two history sources. These results demonstrate the value of paired video-action supervision for improving WAM policies without increasing their deployed sampling budget.

Medical visual question answering (VQA) is a crucial task in clinical AI, yet its evaluation has so far centered almost exclusively on English, limiting its relevance to linguistically diverse patients and clinicians. Recent multilingual medical VQA benchmarks show that large vision-language models (LVLMs) degrade in non-English languages, but lack a fine-grained analysis of how cross-lingual variation affects the distinct capabilities that medical VQA requires. To this end, we construct a multilingual medical VQA benchmark over eight languages, organized into four representative scenarios that isolate the core capabilities medical VQA requires. Evaluating five open- and closed-source LVLMs, we find that cross-lingual degradation is not uniform but highly scenario-dependent. We therefore propose MedVL-XLRepE, a training-free scenario-aware representation engineering method, leveraging LVLMs' superior English medical VQA capability to steer non-English representations toward their English counterparts at inference time. Across three LVLMs and eight languages, MedVL-XLRepE consistently mitigates cross-lingual degradation, with gains of up to 6.33%.

Finite datasets can share the same size and low-order statistics while differing strongly in structural complexity. We connect this dataset complexity to loss-landscape geometry by pairing local label mixing across neighborhood scales with local entropy around trained neural-network solutions. Adapted from the Franz–Parisi construction in spin-glass theory, local entropy measures the effective volume of low-loss, solution-like parameter configurations at each distance from a reference. We estimate it in finite networks using adaptive sequential Monte Carlo. In a controlled synthetic sweep, greater dataset complexity produces a larger decrease in local entropy near the reference. Farther away, its radial derivative becomes weak and nearly common across conditions. Dataset complexity therefore changes where the effective solution volume contracts, rather than making it decrease uniformly faster. Experiments on real image data show the same qualitative trend, with label randomization further amplifying the effect. These results show that dataset structure shapes how low-loss neighborhoods are organized across finite distances from trained solutions.

We introduce MCSI, a two message key exchange we design over the CSIDH class group action, in which each party sends its ephemeral public element under an authenticated encryption keyed by the value the two static keys determine. The design gives implicit mutual authentication, hides the ephemeral element from an eavesdropper, and lets a recipient discard an unauthenticated message after one tag check rather than after an evaluation of the group action, which is four orders of magnitude more expensive. On the analytic side, we prove that our protocol is correct with zero error, and prove three statements in the random oracle model, all reducing to the strong parallelisation problem: indistinguishability of the session key against a passive adversary, confidentiality of the blinded ephemeral element, and integrity of the blinded transport. None uses the decisional group action assumption, which is false for class group actions of non-prime discriminant. We also show that a blinding key cannot come from the session secret it is meant to establish. To instantiate the design we select parameters and show that a prime chosen for elliptic curve discrete logarithms is unusable: for $p = 2^{521}-1$, the NIST P-521 prime, the action admits no efficiently evaluable generator. On the practical side, we build and test the design. We implement the protocol twice, in C and independently in Python, cross check the two, and measure what a session costs in field operations, time and memory. We also audit our code for secret dependent control flow: the field arithmetic and the symmetric layer show none, while the group action leaks the key by construction, and two hundred timings separate two keys whose one-norms differ by five out of 370. Finally, we state what we do not prove, among them security under ephemeral key reveal, forward secrecy of the blinding, and constant time execution.

Automatically analyzing hours-long egocentric video is increasingly essential for progress monitoring, quality control, and safety in logistics, construction, and manufacturing. Yet current pipelines that process short, fixed-size windows with a vision-language model (VLM) are prohibitively expensive because cost scales with the number of model calls. To reduce this cost, prior work proposes triage policies to select which windows merit a VLM invocation. However, these policies either sample uniformly or rank windows using visual features, which ironically requires the video decoding that the budget constraints are meant to avoid. We propose audio-first triage: select windows using the lightest modality, scored before any video frame is decoded, so the approach composes naturally with token compression or quantization. The novelty lies in the objective, not the representation: rather than a per-frame sound-event detector, we train the selector to trigger once per action. This objective shift improves action coverage by 4.0-10.8 percentage points across all evaluated call rates, using frozen AudioSet-pretrained features without domain-specific sound-event labels. Using fewer than half of the available calls, the triage cuts 9-20% of VLM calls at matched coverage on EPIC-KITCHENS-100 (EK-100), surpasses uniform sampling through the mid-range on Ego4D over 247 clips, and outperforms two recent visual keyframe selectors. Code, the reference implementation and every results file this manuscript reads are at https://github.com/masjalayer/PreDecoding-AcousticTriage.

Deep atmospheric convection governs South Asian monsoon variability, yet attempting to learn its latent world model directly from zero-inflated, heavy-tailed precipitation yields suboptimal predictive representations. Continuous atmospheric proxies, such as outgoing longwave radiation (OLR), express this convective organization far more coherently. We address this mismatch with cross-variable proxy learning: M-JEPA, a multiscale Monsoon Joint-Embedding Predictive Architecture, is pretrained on five continuous proxy fields over Lagrangian patches tracking moving convective systems—without rainfall supervision at any point. The resulting frozen representation is transferred to daily precipitation forecasts through a shared decoder trunk featuring parallel probabilistic and deterministic branches. Because rainfall is strictly unobserved during pretraining, downstream skill directly measures the predictive information captured in the latent rollout. A frozen-backbone probing framework with two controls (an identical architecture trained on rainfall alone, and a randomly initialized backbone) attributes the transfer specifically to proxy pretraining: direct rainfall training exhibits $36\%$ higher CRPS error ($7.52$ vs.\ $5.54$\,mm/day). Against the 51-member operational ECMWF ensemble, the transferred model attains a statistically resolved CRPS advantage ($6.81$ vs.\ $6.89$\,mm/day) and higher Brier skill ($+0.05$ vs.\ $-0.04$) using $15.4$M parameters on a single consumer GPU, concentrated at heavy-rain thresholds and fine spatial scales, while the ensemble retains an advantage in neighborhood skill and deterministic references on point metrics. The result provides a competitive monsoon precipitation forecast grounded in intraseasonal dynamics and a diagnostic framework for evaluating transferred atmospheric representations.

Explainable AI (XAI) research has produced a plethora of explanation techniques, yet user studies repeatedly show that available explanations are not effective in practice. We argue that, given the siloed nature of conventional XAI, users are struggling to select the appropriate XAI technique. Viewing XAI through a philosophical lens, we offer a formalization of what we call the selection problem: the systematic failure of XAI interfaces to bridge the gap between a user's natural-language uncertainty and the explanation technique that resolves it. Following a logical premise-conclusion format, we show that conventional interfaces require users to translate their uncertainty into a technique selection, a challenging prerequisite to meet. We also propose a structural solution: a multi-agent LLM orchestration tool that translates the user's query to the proper XAI explanation technique. We provide an example of how this structural solution could be instantiated to address the selection problem.

The 3SUM problem asks, given sets $A,B,C$ of integers, whether there exist $a\in A$ and $b\in B$ whose sum belongs to $C$. In the preprocessed variant with unknown $C$, one preprocesses sets $A$ and $B$, each of size $n$, and subsequently answers a query specified by subsets $A'\subseteq A$, $B'\subseteq B$ and a target set $C'$ of size $O(n)$, by solving the 3SUM instance $(A',B',C')$. Kirkpatrick, Kuszmaul, Mathialagan, and Vassilevska Williams [ICALP 2026] gave the first algorithm with subquadratic space for this problem, achieving $\widetilde{O}(n^{3/2+\epsilon})$ query time using $\widetilde{O}(n^{2-2\epsilon/3})$ space, for every $\epsilon\in[0,1/2]$. Their algorithm employs separate mechanisms for heavy and light targets, and for each heavy target it stores explicitly the list of pairs $(a,b)$ whose sum is congruent to it modulo a random prime $p$ but differs from it. These lists dominate the space bound. We present a unified construction that uses a single mechanism for all queries. Instead of storing these lists of pairs, our algorithm recovers on demand the residue class of the target modulo a random prime by inverting the function $(a,b)\mapsto (a+b\bmod p)$. Additionally, instead of storing the heavy targets, it recomputes at query time a set that covers them. This leads to an improved space bound of $\widetilde{O}(n^{2-\epsilon})$, while maintaining the same query time. Moreover, our construction is the first to achieve subquadratic space while supporting adaptively chosen queries.

Delta-Rule recurrent models maintain a fixed-size state, enabling $O(1)$ inference memory but potentially becoming unstable under extreme-context extrapolation. By tracking RWKV-7 over sequences of up to 100M tokens, we empirically identify a distinct failure pattern: localized norm explosion atop a relatively sparse substrate, rather than global state saturation. Analysis of the recurrent update suggests that persistent decay keeps weakly updated entries small, whereas uneven injections allow a few channels to accumulate extreme values. Motivated by this diagnosis, we propose State Anomaly Neutralization (SANE), which applies adaptive $\tanh$ compression at chunk boundaries while preserving the intra-chunk parallel structure. Within a safe threshold range ($3 \le \alpha \le 5$), SANE matches the baseline on 11 short-context reasoning benchmarks with no statistically significant degradation. After a 100M-token prefix, which exceeds the training length by over $24{,}000\times$, SANE retains functional reasoning ($33.46$–$35.56$) while the baseline encounters numerical overflow. In contrast, overly permissive thresholds ($\alpha \ge 8$) remain numerically stable but lose reasoning capability entirely, showing that numerical stabilization alone does not guarantee functional reasoning and revealing a capacity–stability trade-off in state compression.

This work presents a mathematical and computational framework for computing the dispersion relations of shear horizontal (SH) waves in continuously varying heterogeneous layered structures. The approach isolates the contribution of the heterogeneous layer from the complete dispersion relation, learns this contribution using a physics-informed neural network (PINN) and subsequently incorporates the trained model to determine the complete dispersion relation. The mathematical properties of the discretized problem are investigated, including the singularity of the finite-difference system and the oscillatory behavior of the layer solution, while a generalization-error estimate is established for the PINN approximation. The framework is first tested on a seismological configuration consisting of a heterogeneous sandstone layer over a granite half-space, where exponential heterogeneity is considered with independent variation rates in shear modulus and density. The proposed approach is validated against analytical solutions in special cases, while for general configurations the Haskell matrix method demonstrates convergence toward the continuously varying dispersion relation predicted by the PINN as the number of homogeneous sublayers increases. Parametric studies further confirm consistency with the underlying physics. An important feature of the method is that the heterogeneous-layer equation can be trained independently and reused in multiple settings. To demonstrate this, the same trained network is coupled with piezoelectric and piezomagnetic substrates governed by fundamentally different physical laws, highlighting the potential of the PINN framework as a reusable computational module for dispersion analysis in heterogeneous layered media.

The Erdős-Pósa theorem states that for every integer $k\geq1$, every graph contains either $k$ vertex-disjoint cycles or a set of $\mathcal{O}(k\log k)$ vertices meeting all cycles. This fundamental min-max duality has been extended to numerous settings, including long cycles, $S$-cycles, that is, cycles containing a vertex in a prescribed set $S$, and cycles satisfying various additional constraints. In contrast, much less is known when the packing itself is required to be induced, namely, when distinct cycles are vertex-disjoint and have no edges between them. We prove that long $S$-cycles admit an induced version of the Erdős-Pósa-type duality. More precisely, we show that there exists a polynomial function $f(k,\ell)$ such that for all integers $k\geq1$ and $\ell\geq3$, every graph contains either an induced packing of $k$ $S$-cycles of length at least $\ell$ or a set of at most $f(k,\ell)$ vertices whose closed neighbourhood intersects all $S$-cycles of length at least $\ell$. The proof introduces a new ear-decomposition technique based on fragile ears and yields a polynomial-time algorithm for every fixed $\ell$.

A cognitive architecture is more than the module that reasons: it must also decide how long to think and what deserves the effort. We built a minimal but complete system - a recurrent reasoner with adaptive halting, a homeostatic control field, and a value module - and asked of each part: does this function emerge from gradient descent, or must it be computed? Competence emerges. Stopping appears to emerge too, and to be worth more than everything decidable in advance, but that appearance is instrumentation: payoff at matched mean compute climbs from 0.467 (uniform) through 0.546 (difficulty) to 0.698 (ex-ante value), and the further climb to 0.921 (posterior self-observation) does not survive audit. PonderNet-style halting returns a halting-weighted mixture of hidden states while forced-depth baselines return one, and the language head is trained on the mixture alone; equalizing the readout annihilates the apparent advantage of native execution (residual +0.000 [0.000, 0.000]). Value does not emerge: trained couplings capture zero of a payoff an explicit allocator captures completely (+0.151, routing correlation +0.79), so the second-order decisions that pay must be computed, at least where value is orthogonal to content, as here by construction. On a frozen LLM actuator the same instruments show self-consistency voting to be a measured bound (+0.0236 [+0.0150, +0.0326]) and inter-sample agreement nearly worthless as a stopping signal, its mass concentrating on wrong answers. Every null we assert carries a mechanism and a positive control, and the protocol is part of the contribution. Executing our own falsifiable prediction, value under commitment pays +0.1312 [+0.1124, +0.1502] in a cliff-cost family, some seven times the smooth-family estimate - not because the cliff shifts information ex ante, but because it multiplies the attainable range fivefold (5.1x [3.4, 8.2]).

This data article describes a multimodal collection of scanned examination answers paired with expert-designed Outcome-Based Education (OBE) grading metadata. The collection contains 485 answer submissions from 415 consenting students at four academic institutions. Eight faculty contributors supplied examination materials covering nine subjects and 12 distinct question templates. Each answer-level item links a scanned PDF to a randomized identifier, subject label, question, model answer, criterion definitions, performance-level descriptions, criterion marks, and a total mark. The 12 rubrics contain 47 criteria in total. The scans retain realistic academic content, including handwriting, printed text, equations, tables, code, figures, sketches, and diagrams. CamScanner, Adobe Scan, and conventional scanners contributed variation in illumination, contrast, orientation, compression, and resolution. Diverse handwriting, crossed-out work, revised calculations, and inserted corrections add further visual variability for robustness and generalization studies. Preparation involved heterogeneous-source consolidation, label and text standardization, score validation, identifier randomization, filename randomization, and JSON-to-PDF integrity checks. An answer-level audit confirmed 485 unique identifiers, 485 unique PDF filenames, agreement between each total mark and its criterion-mark sum, and scores within the applicable rubric maximum. The data can support rubric-aware automated evaluation, multimodal document understanding, criterion-level feedback, score prediction, and privacy-aware OBE assessment research. Access is restricted to research use and is available from the corresponding author upon reasonable request.

Smart contracts power blockchain applications such as DeFi and NFTs. However, once deployed, they cannot be modified. Even minor bugs can result in significant financial losses. Current AI-based repair methods rely on linear reasoning, which leads to the accumulation of errors and unreliable patches. Our method combines document parsing, static analysis, and Tree of Thoughts reasoning. We first convert audit reports into structured data. Then we use Slither to locate the exact vulnerable code. Our three-step framework explores multiple repair paths simultaneously, evaluates options, and eliminates poor choices. Finally, we verify patches through compilation and manual checks. We test our method on 50 real vulnerabilities from Code4Rena. Our method achieves a 62% single success rate and an 84% top-3 success rate, outperforming ContractTinker by 12 and 6 percentage points, respectively. We also increase the proportion of fully effective patches to 44%, while reducing defective patches from 38% to 22% and invalid patches from 10% to 4%. This approach overcomes the limitations of linear reasoning and makes smart contract repair more accurate and practical.

3D Gaussian Splatting (3DGS) achieves state-of-the-art rendering quality at real-time speeds but suffers from "model bloat" - a large number of redundant, low-opacity Gaussians that inflate memory usage and training costs. This inefficiency stems from the standard "densify-then-prune" paradigm, which expands the model aggressively before relying on pruning to achieve compactness. To mitigate this problem, we present an efficient training framework that builds an intrinsically compact representation, replacing the conventional densify-then-prune cycle. Our method leverages a synergistic design: an L2 reconstruction loss to provide error-proportional gradients that stabilize optimization, and a novel Polarized Opacity Prior (POP) to actively manage the Gaussian population. POP steers informative primitives toward full opacity and uninformative ones toward transparency, enabling natural pruning and accelerating rendering through Early Ray Termination. Experiments on three public datasets demonstrate that our approach consistently achieves accelerated 3DGS training with significantly fewer Gaussians while maintaining comparable visual reconstruction quality. These results show that the proposed framework provides a simple and effective path toward fast and inherently compact 3DGS training.

Lifting 3D hand poses from 2D monocular representations remains challenging due to the limited availability of large-scale, diverse 3D-annotated hand datasets, in contrast to the abundance of human body motion data. We address this limitation by transferring motion representations learned from large body pose corpora to the hand domain. We introduce TransHands, a backbone-agnostic transfer learning framework that enables pre-trained human motion encoders to be effectively adapted for 3D hand pose estimation from 2D pose inputs. Rather than training hand-specific biomechanical models from scratch, TransHands combines a two-stage training and fine-tuning strategy with a lightweight hand-specific input adaptation module that aligns hand kinematics with the representation space learned for full-body motion. We evaluate TransHands across four state-of-the-art motion modeling architectures, including transformer-based, graph-based, and frequency- domain models. Results demonstrate that motion priors learned from body pose data transfer consistently across architectures, yielding consistent accuracy gains, strong cross-domain generalization, particularly in challenging egocentric settings, and applicability for downstream tasks in real-world contexts.

Extracting skills from past successes is critical for the efficient evolution of Large Language Model (LLM) agents. Prevailing agent self-evolution paradigms typically rely on a core assumption: equipping LLMs with skill memories derived from successful trajectories will monotonically improve their problem-solving capabilities. However, probe analyses reveal that extracting skills solely from successful trajectories traps the model in a Skill Imitation Trap. For tasks that resemble past successes but require different tools, retrieving more skills paradoxically increases the model's confidence in wrong tool calls—procedure skills raise the wrong-tool margin by $47\%$ over a memory-free baseline. To overcome this limitation, we propose Boundary-Aware Skill Memory (BASM), which augments each skill with explicit boundary fields—applicability conditions, risk cues, avoidance rules, and recovery notes. These fields transform each retrieved skill from an unconditional action template into state-conditioned guidance: the agent applies the skill when its conditions hold, suppresses inapplicable tool calls when they do not, and issues targeted repairs when execution fails. Across three agent benchmarks and four model scales, BASM consistently outperforms success-distilled skill-memory baselines: it improves task success rate by up to $23.8\%$ on AppWorld, accuracy by up to $5.0\%$ on BFCL, and reduces attack success rate by $4.6\%$ on AgentDojo, while simultaneously reducing average AppWorld steps by up to $6.6\%$ relative to the memory-free baseline.

Interactive PDF form fields are often absent from documents that visually resemble forms, leaving users unable to enter data without printing or external editing tools. Detecting the missing widgets is difficult because a field may be indicated by several overlapping cues, born-digital PDFs expose useful but incomplete drawing structure, and dense pages can contain hundreds of fields. We introduce AcroMELD (AcroForm Multi-source Evidence Linking Decoder), a 39.4M-parameter detector that combines a high-resolution visual transformer with label-free PDF primitives. Its 896-query set comprises 384 visual proposals, 384 structure-seeded proposals, and 128 learned recovery queries. Four graph-set layers exchange information over geometry-biased sparse neighborhoods and cross-attend to PDF structure. A learned same-field relation links co-referent candidates, while a localization-quality head is trained on the containment-aware overlap used by the downstream recovery decision. We define a hash-bound evaluation protocol with disjoint development, calibration, internal-test, and quarantined external-holdout roles. The sealed, single-seed candidate reaches native containment micro-$F_1$ 0.9344 on the internal test and 0.8477 on the one-shot external holdout (95% PDF-cluster bootstrap interval [0.8339, 0.8605]). This passes the registered historical FFGBT-v8 reference by 0.0186 absolute $F_1$. Under the stricter external adapter, however, performance is 0.7786 IoU-$0.5$ $F_1$ and 0.2900 COCO mAP, below a locally evaluated CommonForms-L reference; the signature class receives no prediction at the selected threshold. Thus the result supports the registered operational gate while exposing substantial domain and rare-class limitations.

Speech-guided referring video object segmentation aims to recover the mask tracks of objects specified by a spoken motion description. Here, speech carries a linguistic instruction rather than acoustic evidence from a sounding object, so a solution must connect speech recognition, motion-centric temporal grounding, mask tracking, and explicit no-target handling. We introduce Speech2MaskTrack, our approach for the MeViS-Audio track of the 8th LSVOS Challenge. Speech2MaskTrack transcribes the spoken query and compiles it into structured constraints over category, count, direction, interaction role, and temporal phase. SAM3.1 enumerates multiple instance tracks, which TRACE ranks using complete-trajectory motion and relation evidence. A frozen lexical presence gate may suppress the ranked SAM3.1 base prediction. When the gate predicts that a target is present, an available full-expression-conditioned SaSaSa2VA track replaces the SAM3.1 mask. Only outputs that remain empty enter GPT-assisted recovery, which invokes SaSaSa2VA again under query- and mask-level verification. Speech2MaskTrack achieved second place in the official challenge ranking.

Bengali is the seventh-most-spoken language globally, yet LLM safety evaluation remains overwhelmingly English-centric. We introduce BanglaSafe, a benchmark of 879 Bengali prompts combining 309 natively authored prompts with 570 expert-reviewed prompts, spanning 17 culturally grounded harm categories and five prompting conditions that vary language, writing style, and authority framing. Evaluating 18 frontier LLMs, we find that over half of all responses are unsafe or partially unsafe (53.6%) while 14.7% contains strictly harmful content, and that the strongest observed effect is not the switch from English to Bengali but the choice of writing style within Bengali: the same harmful request phrased as a formal newspaper investigation succeeds 17 percentage points more often than the same request phrased as a casual message, with no adversarial engineering involved. We further show that existing safety classifiers struggle to reliably evaluate Bengali content, with even frontier models failing on nearly half of all cases.

Near a smooth data manifold, one tangent space summarizes local geometry. At a branch point, the corresponding first-order object is instead a measure over tangent directions, whose normalized masses record the local share of each branch under the chosen data measure. We ask whether a score field at one noise level determines this weighted tangent geometry when the branch center and homogeneity degree $d$ are unknown. In this tangent-measure model, $d$ is the local measure dimension. Gaussian smoothing of a homogeneous tangent measure satisfies an Ornstein–Uhlenbeck eigenfunction equation. Its weak form turns score values—without score derivatives—into a linear system for the center and homogeneity degree, with an explicit rank condition and perturbation bound. After this calibration, the tangential score on one sphere is the spherical log-gradient of a scalar Gaussian–cone transform. Integration recovers that transform up to scale, and all its spherical-harmonic multipliers are positive. Thus one exact shell identifies the normalized angular measure in every ambient dimension $D\geq2$. For at most $K$ positive rays, moments through degree $2K-1$ constructively recover count, directions, and weights in arbitrary dimension. Any fixed observation scheme needs at least $KD-1$ scalar tangential components. In the plane, degree $K$ is both sufficient and necessary, and we give quantitative finite-query certificates. For finite planar $C^{1,\beta}$ branches with positive $C^{0,\beta}$ densities, we prove $O(\sigma^\beta)$ convergence from the finite-noise score to its tangent model. In controlled experiments, 50k-step training lowers validation normalized-score error across four geometries yet raises angular-moment error, separating ordinary score fit from geometry recovery.

Large Language Models (LLMs) demonstrate remarkable problem-solving capabilities when guided by Chain-of-Thought (CoT) prompting, yet the internal mechanisms underlying these improvements remain poorly understood. In this work, we investigate where CoT-related causal effects emerge across the generated reasoning trajectory and which attention heads carry signals that contribute to final-answer computation. Because CoT reasoning unfolds over multiple generated tokens, standard activation patching at a single static token position is insufficient to characterize these temporally distributed effects. To address this limitation, we introduce a sequential activation patching framework that traces CoT-conditioned attention-head activations across token positions and aggregates their effects using Part-of-Speech-guided analysis. We further introduce Sequential Multi-Head Patching to evaluate the joint contribution of distributed head sets, together with cross-question and random activation controls. Targeted zero-ablation experiments show that the identified heads are functionally important for successful answer generation and affect several overlapping mechanisms, including reasoning-trajectory maintenance, answer anchoring, exemplar-target separation, and numerical generation. Overall, our results provide evidence for distributed reasoning-support sub-circuits associated with CoT-conditioned computation.

We audit measurement variability for 3 native tool-calling endpoints across 2 providers on the official BFCL multiple and parallel categories, using matched AST grading. At temperature 0, reruns are nearly deterministic across Groq endpoints and a thinking-enabled Gemini setting: ever-flip fractions are 0.7%, 2.0%, and 2.7%, with mean run correlations of 0.997, 0.966, and 0.961. Semantics-preserving prompt perturbations create the larger floor on all endpoints, with median perturbation paired SDs 11x to 58x larger than rerun paired SDs. The failure character also shifts: malformed-output failures account for 30%, 7%, and <1% of task failures, so marginal accuracy hides not only stability but also failure mode.

Emotion-aware artistic image generation requires a model to satisfy semantic content, artistic style, and target emotion simultaneously. The key challenge is that artistic captions conflate these axes into underspecified free-form text, making fine-grained visual attributes such as brushwork, composition, and tonal atmosphere difficult to ground concretely. We present ReART, a reference-guided retrieval and refinement framework. Our method decomposes test captions and each image annotation in the EmoArt database into structured visual fields, and performs field-wise retrieval over subject, layout, brush-line, and tone-mood dimensions to retrieve role-specific visual references that supply the perceptual detail text alone cannot convey; these references are used alongside a structured prompt for initial synthesis. For samples where any Attribute Alignment Score (AAS) axis falls below threshold, an AAS-driven refinement loop diagnoses failures, constructs constrained repair plans specifying elements to keep, errors to fix, and operations to avoid, routes references by correction purpose, and performs controlled editing under structural preservation constraints. Our system ranks 2nd in Track 1 of the AffectiveArt 2026 Grand Challenge, achieving a perfect AAS of 1.00 and an overall score of 0.78. Code is available at https://github.com/oceanflowlab/ReART.git.

Graph-of-Convex-Sets (GCS)-based trajectory optimization represents collision-free regions in configuration space as a finite collection of convex sets and directly performs collision-free trajectory planning over these sets, substantially simplifying the planning process. However, existing GCS-based trajectory planning methods generally assume sufficient connectivity among the convex regions and do not explicitly address cases in which the start and goal regions belong to different connected components of the initial GCS map. To address this limitation, we propose GCS-Bridging, which reconnects disconnected convex regions through collision-free point paths followed by convex region inflation, thereby recovering the feasibility of otherwise disconnected GCS planning problems. Extensive simulations across multiple IRIS-related algorithms and scenarios demonstrate that GCS-Bridging restores missing start-to-goal connectivity in the initial GCS map with a 99.8% success rate. In addition, a hardware experiment on a single-arm Franka platform in a real-world scenario with initially disconnected start and goal regions validates the effectiveness of the proposed method in practical motion planning. Project website: https://zhouxk1997.github.io/GCS_Bridging/

Parameter estimation is a central task in data-driven learning of dynamical systems. It aims to recover the underlying physical parameters from observed time-series data, thereby providing interpretable insights into the physical mechanisms governing the system. Gradient/derivative matching methods based on Gaussian process provide an efficient way to perform parameter estimation. Those methods avoid repeated numerical integration and enforce local derivative consistency. However, such local matching may result in global inconsistency with the governing flow map, particularly under scarce and noisy observations. To address this limitation, we propose a framework based on Gaussian process learning with flow map refinement (GPL-FMR), a two-stage parameter estimation framework. The first stage is based on Gaussian process learning algorithm and the posterior obtained from which is transferred as an informative prior to the second stage based on flow-map refinement. The second stage further improves the parameter estimation via optimisation based on global dynamical constraints. We demonstrate and analyse its performance on multiple numerical examples, including the Van der Pol oscillator, the Lotka-Volterra model, and the Lorenz-63 system. The results show that the proposed framework consistently improves parameter estimation accuracy, particularly under scarce and noisy observations.

The application of Large Language Models (LLMs) to diagnostic decision-making has garnered growing interest. However, existing benchmarks largely focus on textual reasoning or isolated visual question-answering (VQA) tasks, lacking holistic integration of clinical narratives and medical imaging, and thus failing to assess the multimodal diagnostic synthesis capability central to expert clinical judgment. To bridge this gap, we introduce MedReaMM, a benchmark specifically designed to evaluate models' ability to synthesize heterogeneous clinical evidence consisting of detailed patient histories alongside multiple medical images into accurate differential diagnoses under a complete-information paradigm. Constructed from case reports sourced from top-tier medical journals and curated clinical case databases, MedReaMM comprises 625 expert-validated cases with an average of 2.79 medical images per case and a total of 1,042 standardized diagnoses annotated with ICD-11 codes. These cases predominantly represent rare, atypical, or multi-system presentations that demand expert-level evidence integration beyond routine pattern recognition. We evaluate 23 Large Multimodal Models (LMMs) and find that most achieve diagnostic accuracy scores below 50%, underscoring a substantial gap in multimodal diagnostic synthesis capability. Further analysis reveals that medical knowledge proficiency, medical image understanding, and evidence integration are all highly correlated with diagnostic performance.

Optimizer-state quantization is commonly designed for Adam's dense, parameter-aligned first- and second-moment arrays. This abstraction breaks for memory-efficient optimizers, whose states may be factored, confidence-modulated, or maintained in a projected space, so similar reconstruction error can produce different update error. We formulate optimizer-state quantization as a joint problem over representation, topology, and update semantics. We then introduce Adaptive Log-Space (AL) quantization for non-negative states. AL fits each block's observed nonzero logarithmic interval and reserves a separate code for exact zero, enforcing $q = 0 \Leftrightarrow x = 0$; signed momentum and state precision remain independently selectable. Controlled probes show that adaptive ranges reduce update error and temporal drift, exact-zero reservation preserves dormant states, and state topology constrains useful block granularity. End-to-end language-model training evaluates the resulting policy across dense, factored, confidence, and projected optimizer states. On TinyLlama-1.1B, AL8 with uniform 8-bit momentum reaches 72.90 perplexity versus 73.54 for bitsandbytes 8-bit AdamW, with comparable optimizer-state storage and higher throughput. CAME matches reference-level final perplexity across three seeds when its non-negative states use AL16, while a semantic grouping-and-protection policy closes most of quantized Adafactor's 100K-step late-loss gap. These results make state topology and update semantics first-class design constraints for optimizer quantization.